Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Algorithmic guided screening of drug combinations ofarbitrary size for activity against cancer cells
Ralph G. Zinner,1 Brittany L. Barrett,1
Elmira Popova,4 Paul Damien,5 Andrei Y. Volgin,2
Juri G. Gelovani,2 Reuben Lotan,1 Hai T. Tran,1
Claudio Pisano,6 Gordon B. Mills,3 Li Mao,1
Waun K. Hong,1 Scott M. Lippman,1
and John H. Miller7
Departments of 1Thoracic/Head and Neck Medical Oncology,2Experimental Imaging, and 3Systems Biology, The University ofTexas M. D. Anderson Cancer Center, Houston, Texas;4Department of Operations Research and Industrial Engineeringand 5Red McCombs School of Business, The University ofTexas, Austin, Texas; 6Sigma-Tau Pharmaceuticals, Inc., Rome,Italy; and 7The Santa Fe Institute, Santa Fe, New Mexico
AbstractThe standard treatment for most advanced cancers ismultidrug therapy. Unfortunately, combinations in theclinic often do not perform as predicted. Therefore, tocomplement identifying rational drug combinations basedon biological assumptions, we hypothesized that afunctional screen of drug combinations, without limitson combination sizes, will aid the identification of effec-tive drug cocktails. Given the myriad possible cocktailsand inspired by examples of search algorithms in diversefields outside of medicine, we developed a novel, efficientsearch strategy called Medicinal Algorithmic Combinato-rial Screen (MACS). Such algorithms work by enriching forthe fitness of cocktails, as defined by specific attributesthrough successive generations. Because assessment ofsynergy was not feasible, we developed a novel alterna-tive fitness function based on the level of inhibition andthe number of drugs. Using a WST-1 assay on the A549cell line, through MACS, we screened 72 combinations ofarbitrary size formed from a 19-drug pool across fourgenerations. Fenretinide, suberoylanilide hydroxamic acid,and bortezomib (FSB) was the fittest. FSB performed up
to 4.18 SD above the mean of a random set of cocktails or‘‘too well’’ to have been found by chance, supporting theutility of the MACS strategy. Validation studies showedFSB was inhibitory in all 7 other NSCLC cell lines tested. Itwas also synergistic in A549, the one cell line in whichthis was evaluated. These results suggest that whenguided by MACS, screening larger drug combinations maybe feasible as a first step in combination drug discoveryin a relatively small number of experiments. [Mol CancerTher 2009;8(3):521–32]
IntroductionCombining chemotherapeutic agents is an establishedway to improve on single-drug efficacy, and small (two-to three-drug) combinations are now the standard oftreatment for most metastatic cancers (1). Recent successeswith larger multiple-agent regimens in the clinic (1, 2)encourage a continued search for improved combinationchemotherapies; however, only a few of the virtuallyinfinite possible combinations have been evaluated to date(3, 4) due to concerns about the potential toxicity of thecombinations and the daunting logistics of testing allfeasible combinations of available drugs (5).The improved tolerability of new, potentially effective
molecular targeted agents has encouraged combination ofthese agents with standard chemotherapy regimens and,more recently, with each other in early-phase clinicaltrials (1, 2, 6). In a recent review, it was observed thatthere is ‘‘tremendous’’ potential for improved therapy ofcancer from combinations of molecular targeted agentsdesigned to modulate different aspects of the same ordifferent targets (7). Traditionally, strategies of combinationdrug development have been based on insights into non–cross-resistant molecular mechanisms of action, evidenceof synergy, preclinical or clinical insights into nonoverlap-ping toxicities, and, more recently, insights into theireffects on signaling pathways and host-tumor interactions(8, 9).As a complement to these methods, we investigated the
utility of a functional, laboratory-based screening ofcombinations. Laboratory-based screens have long beenused to identify active single agents, and one companyexplored doublets using synergy (10). However, to ourknowledge, there are no ongoing efforts to screen largespaces of potential combinations of more than two drugs.Extending a screen to larger cocktails may be useful giventhe time required to compile clinical evidence. However,in vitro assessments of any given cocktail become morechallenging as drug combinations use larger numbers ofagents. Moreover, even with highly efficient assessmentsof individual combinations, only a small fraction of thetotal possible cocktails can be evaluated given that they are
Received 9/29/08; revised 1/9/09; accepted 1/15/09; publishedOnlineFirst 3/10/09.
Grant support: Department of Defense BESCT Lung Cancer ProgramDAMD17-01-1-0689 (W.K. Hong).
The costs of publication of this article were defrayed in part by thepayment of page charges. This article must therefore be hereby markedadvertisement in accordance with 18 U.S.C. Section 1734 solely toindicate this fact.
Requests for reprints: Ralph Zinner, Department of Thoracic/Head andNeck Medical Oncology, Unit 432, The University of Texas M. D.Anderson Cancer Center, 1515 Holcombe Boulevard, Houston, TX77030. Phone: 713-792-6363; Fax: 713-792-1220.E-mail: [email protected]
Copyright C 2009 American Association for Cancer Research.
doi:10.1158/1535-7163.MCT-08-0937
521
Mol Cancer Ther 2009;8(3). March 2009
Published Online First on March 10, 2009 as 10.1158/1535-7163.MCT-08-0937
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
virtually countless. We made an effort to meet thesechallenges on two levels. First, we developed a rapidmethod, or fitness function alternative to synergy, to speedevaluation of individual cocktails. Second, we used a searchalgorithm intended to sequentially direct the screen towardincreasingly more promising sets of drug combinations.Fitness is a term used in search algorithms adapted
from evolutionary theory. In biology, higher fitnessdescribes an individual’s greater capacity to reproduce.In search algorithms, it indicates a more optimal solution;what defines fitness is determined by the investigator andthe problem that he or she is trying to solve. Thesealgorithms guide a process through a finite list ofoperations intended to maximize fitness. Such algorithmshave been extensively used outside medicine in applica-tions such as data fitting, scheduling, trend spotting, andbudgeting (11–13). They have been evaluated in medicine
in breast cancer diagnosis (14) and in an antiviral in vitromodel assessing a large space of dose permutations of afew cocktails (15). Here, a well-established strategy isapplied to a novel setting: the discovery of drug cocktails.The algorithm works through a stepwise process ofenriching successive pools of cocktails for increasedfitness. It is this application we designate MedicinalAlgorithmic Combinatorial Screen (MACS).For MACS to be meaningful, fitness needs to predict
a value of interest, and to be feasible, the fitness functionneeds to be efficient. We intended fitness to predictcocktail efficacy. Synergy is an established fitness functionused for this purpose. Indeed, synergy was used byCombitoRx in evaluating doublets (10). However, synergycan become exponentially more labor-intensive witheach drug added to a given cocktail. As an example,when six doses per drug are used, as done by CombitoRx,
Figure 1. A, a model hill climb. Imagine an eight-drug pool with the drugs labeled ‘‘a’’ through ‘‘h.’’In the initial generation (generation 0), 9 of 256possible combinations are randomly formed. Thesecombinations are, in turn, tested in cell culture. Inthis example, the combination identified as fittest isshown as ‘‘bdeg.’’ Note that we are not describingthe fitness function here; different fitness functionscan alter the fitness landscape and which among thenine would be determined to be fittest therebyaltering the ultimate outcome. However, the rules ofthe hill climb will not be affected by the fitnessfunction. Generation 1 is formed from the eight one-mutant combinations (also known as ‘‘nearestneighbors’’) of ‘‘bdeg’’ plus the parent. To formthis generation, each of the eight drugs is added toor subtracted from the fittest combination in turn.Thus, because drug ‘‘a’’ is not present in the parentcombination ‘‘bdeg,’’ it is added to form ‘‘abdeg.’’Next, drug ‘‘b’’ is checked. Because it is alreadypresent in the parent, ‘‘b’’ is subtracted, resulting inthe combination ‘‘deg.’’ This is repeated until alleight drugs are checked and generation 1 is formed,which is then tested. If ‘‘deg’’ were identified as thefittest combination, its eight nearest neighborswould then be determined, and so on. A local peakor maximum on the landscape is reached if none ofthe progeny are superior to the parent. B, a schemeoutlining general steps characteristic of any MACS.
MACS to Identify Therapeutic Drug Combinations522
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
cocktails containing two, three, four, or five drugsgenerate 36, 216, 1,296, and 7,776 possible dose permuta-tions, respectively. Although substantially smaller numb-ers of permutations could be managed by fixed dosingratios or other ways of sampling, synergy may nonethe-less remain daunting. Thus, we developed an alternativefitness function that required one assay per cocktail toimprove the efficiency on the level of the individualcocktail. It was a composite of inhibition at a fixed doseof each drug and the number of drugs. Through the useof this novel fitness function, we lost information abouta vast number of potentially informative dose combina-tions, but it enabled us to screen hundreds of uniquecocktails per week. Other fitness functions could incorpo-rate drug cost, anticipated toxicities, molecular insight,dose sequence, alternative dosing schemes, or any othercharacteristics of interest.As with fitness functions that assess individual cocktails,
there are countless possible search algorithms to guidescreens of large sets of cocktails with different onessuited to varied combinatorial landscapes (SupplementaryFig. S1).8 We present a simple example of such analgorithm called hill climbing (Fig. 1A; SupplementaryFig. S1).8 This example does not represent actual data but
rather illustrates a method. As can be seen in the figure,one can perform a MACS by sampling a relatively smallsubset of the possible cocktails (Supplementary Fig. S1).8
Although the rules can vary among MACSs, they all entaila cycling between experimental evaluation of drug cock-tails and a rule-based creation of the next generation(Fig. 1B). These rules involve selection and alteration. Theselection function directs the choice of a subset of thecocktails tested, and in the case of the hill climb, the fittestone. However, it could choose more than one cocktail andeven incorporate a stochastic element analogous to naturalevolution in which fitness improves the odds of survivalbut offers no guarantees (as presented previously).9,10 Thealteration function can create all of the nearest neighborsof the selected cocktail(s), analogous to mutations as in thehill climb, or could combine pieces of different selectedcocktails analogous to recombination.Because any search of a large enough potential space
will, for practical reasons, be confined to a small fractionof the total possible cocktails, it is likely that many ofthe fittest combinations will never be tested even whena search algorithm guides the process. The MACS needsonly to find combinations that have fitness levels that are
8 Supplementary material for this article is available at Molecular CancerTherapeutics Online (http://mct.aacrjournals.org/).
Table 1. The average activity of the single-agent controls across generations 0 to 3 hill climb
Agent (designation) Drug class/mechanism of action IC10 dose (Amol/L) 1� dose 2� dose 4� dose Rank 1� Rank 4�
Anisomycin Protein synthesis inhibitor/activatesJNK, p38 MAPK, and others
0.015 0.90 0.72 0.59 12 12
ATRA Retinoid: differentiation 15 0.87 0.74 0.63 8 16
Bortezomib 26S proteasome inhibitor 0.005 0.97 0.45 0.37 17 4
CD437 Retinoid: proapoptotic 0.3 0.86 0.61 0.34 7 2Cisplatin Alkylating agent 5 0.85 0.72 0.51 5 9Decitabine DNA methylation inhibitor 2 0.88 0.84 0.78 10 17Deguelin Akt inhibitor 12 0.79 0.64 0.39 2 5Fenretinide Retinoid: proapoptotic 5 0.86 0.66 0.19 6 1
Gemcitabine Antimetabolite 0.004 1.06 0.96 1.03 18 19Imatinib TKI of BCR-ABL, PDGF, c-kit 3 1.07 0.78 0.85 19 18Indirubin PKI*: GSK-3b, CDK5 1 0.92 0.69 0.51 15 10LY294002 PI3K inhibitor 2 0.90 0.79 0.53 11 11MX3350-1 Retinoid: proapoptotic 0.5 0.92 0.68 0.49 14 8PD-168393 TKI: EGFR 5 0.85 0.63 0.47 4 7Rapamycin mTOR inhibitor 6 0.76 0.66 0.62 1 15SAHA Histone deacetylase inhibitor 2 0.88 0.50 0.37 9 3
SCH66336 FTI 5 0.90 0.78 0.59 13 13SP600125 JNK1, -2, and -3 inhibitor 5 0.83 0.66 0.45 3 6ST1926 Adamantyl retinoid: proapoptotic 0.05 0.95 0.90 0.60 16 14
NOTE: The three drugs in FSB and ATRA are in boldface. The 1� dose was the dose used in the MACSs. The levels of inhibition at the 1� and 4� doses areranked and represent the average value across all generations (see Supplementary Fig. S2 for charts of each agent across the generations).
Abbreviations: ATRA, all-trans retinoic acid; SAHA, suberoylanilide hydroxamic acid; PKI, protein kinase inhibitor; TKI, tyrosine kinase inhibitor;PDGF, platelet-derived growth factor; EGFR, epithelial growth factor receptor; FTI, farnesyl transferase inhibitors; JNK, c-jun NH2-terminal kinase; CDK5,cyclin-dependent kinase 5; MAPK, mitogen-activated protein kinase; PI3K, phosphatidylinositol 3-kinase; mTOR, mammalian target of rapamycin.
9 Zinner RG, Barrett BL, Volgin AY, Gelovani JG, Huang J, Tran HT, MillsGB, Hong WK, Fu Y, Mao L. Proc Am Soc Clin Oncol, 2005; abstract 2079.10 Miller JH, Zinner RG, Barrett BL. Directed discovery of novel drugcocktails. Sante Fe Institute Website 2005; 07-031, http://www.santafe.edu/research/publications/wpabstract/200507031.
Molecular Cancer Therapeutics 523
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
‘‘good enough’’ more efficiently than a random screen tobe useful.In this proof-of-principle study, we developed a novel
fitness function to evaluate cocktails and tested a number ofdifferent kinds of MACSs operating on combinationsall derived from the same pool of 19 available anticancerdrugs in an in vitro non–small-cell lung cancer (NSCLC)model. Based on data from earlier MACSs showing that alocal search method such as a hill climb would be efficient(as compared with the more complex algorithms weevaluated earlier),9,10 here we tested a hill climb-MACS.We then ran validation assays on the fittest identifiedcocktail in line with the general strategy in most drugdevelopment screens.
Materials andMethodsUnlike the model (Fig. 1A), the actual laboratory experi-ments composing the hill climb entailed 19 real agentscreating a combinatorial space of 219 or 524,288 possiblecocktails. Eighteen cocktails were randomly formed for itsgeneration 0 (as previously described9,10) and subsequentgenerations were created based on the assay results and therules of the hill climb. The assay systems and conditions aswell as fitness function were held constant throughout thehill climb-MACS.
Fitness FunctionOur fitness function permitted one assay per cocktail to
be done in triplicate. It was composed of a measured assay
value representing inhibition and a nonmeasured value, afactor of the number of drugs in the cocktail. Starting at 0,0.1 was subtracted for each 10% of cells remainingcompared with the no-treatment controls. Thus, 70%inhibition yields 30% of the no treatment control value, or0.3 to be subtracted. A ‘‘penalty’’ of 0.1 was assigned foreach drug composing the cocktail based on the observationthat each drug dosed at IC10 added to a cocktail increasedinhibition by an average of 10% (data not shown). Thispenalty was designed to control for cocktail size to allowfitness to be determined by the cocktail composition.Formally, fitness was �K(s) � 0.1|s|, where K(s) is the
proportion of cells surviving relative to the no-treatmentcontrol and |s| was the number of drugs in the cocktail.Fitness values thus can range from �0.1 (one drug kills allcells) to �2.9 (all 19 drugs with no inhibition). Even lowerfitness values would be possible with growth stimulationrelative to the untreated controls.
Selected AgentsNineteen agents, 16 molecularly targeted and 3 chemo-
therapy, were used based on their availability, affordabil-ity, and ease of preparation and storage and the authors’experience in their use (Table 1; Supplementary Methods).Selected agents were required to inhibit A549 growth by atleast 10% (IC10) under the conditions used in the MACS(data not shown). A549 was chosen because it growsrapidly under uncomplicated laboratory conditions. Drugsrequiring excessive concentrations of DMSO were excluded
Figure 2. A, the performance of the hill climb. B, the set of cocktails composing the fittest cocktail and its nearest neighbors (or progeny) resulting froma local search compared with a set of randomly formed cocktails (described elsewhere).9,10A and B, the x-axis indicates the generation number, and in they-axis, higher values indicate greater fitness level [not the inhibitory, K(s)]. A, the hill climb arrived at FSB by generation 2. Generation 3 includes FSB andits 19 nearest neighbors, which can be compared with the 18 randomly formed cocktails of generation 0. B, in blue is a set of 30 randomly formedcocktails; in red are FSB (the fittest cocktail) and its 19 nearest neighbors. This latter set is the same as the cocktails in generation 2 of the hill climb A,although the values in the hill climb represent a repeat assay of these cocktails. C, the average inhibitory values of the single-agent controls per generationare shown. The 1� values are the doses used in the MACSs. Note that the inhibitory values [or K(s)], not the fitness, are indicated along the y axis. Theaverage values across all their respective generations are shown at the far right. For the individual values of each single-agent control across thegenerations, see Supplementary Fig. S3. D,multiple comparisons of the average fitness values and respective 95% confidence intervals of the generationsfrom the hill climb.
MACS to Identify Therapeutic Drug Combinations524
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
given the independent ability of DMSO to inhibit A549 atdoses of 1.7% by volume (data not shown). Drugs dissolvedin ethanol were excluded because ethanol potentiatedinhibition of A549 by DMSO.
Cell LinesEight human NSCLC cell lines were used to test the
activity of the lead drug combination. A549, H226, Calu-3,H441, H1975, and HCC827 were purchased from AmericanType Culture Collection. HCC2279 was a gift from JohnMinna (Southwestern Medical School, Dallas, TX), and PC-14 was purchased from Riken BioResource Center. A549,H226, and Calu-3 cells were incubated in DMEM/F-12(Invitrogen Corp.) containing 10% fetal bovine serum and1% penicillin/streptomycin. H441, H1975, HCC827,HCC2279, and PC-14 cells were incubated in RPMI 1640plus 10% fetal bovine serum.Cells were isolated from BD Falcon 75-cm2 tissue
culture–treated flasks after a rinse with Cellgro PBS(Mediatech, Inc.) and the addition of Gibco trypsin-EDTA(0.05% trypsin, with EDTAx4Na) from Invitrogen. The cellswere then plated at a volume of 50 AL/well in 96-wellMicroWell Nunclon-D plates (Nalge Nunc International).Plated cells were then stored at 37jC in a humidifiedatmosphere containing 5% CO2 until the drugs were added.The numbers of cells per well were as follows: A549, 1,500;H226, 3,000; PC-14, 2,500; HCC2279, 4,000; HCC827, 7,000;H441, 2,700; H1975, 6,000; and Calu-3, 15,000. All cell lineswere incubated on Nunc 167008 96-well plates (FisherScientific International, Inc.).
Preparation of the Drug CombinationsDrugs for a given combination were mixed in a 96-well
2-mL sterile polypropylene block from Denville Scientific.All combinations were mixed at four times the volumeneeded for one well to allow for triplicate experimentsand pipetting error. The final volume of DMSO (0.66%)in all wells, including the no treatment controls, wasbased on the maximum possible DMSO dose derivedfrom the one cocktail containing all 19 drugs. Becauseeach cocktail had its own individual DMSO volume, weadded the unique additional DMSO individually to eachcocktail, yielding 0.66% of a final volume of 250 AL.Medium was added to obtain a volume of 200 AL, whichwas in turn added to a well already containing the cellsand media plated immediately before occupying 50 AL,yielding 250 AL.Drug DosesWe used the single-agent IC10 dose (1� dose) regardless
of the size of the combination because we anticipated thata higher IC value would result in maximal inhibition frommany of the small cocktails, thereby reducing our abilityto compare cocktails. All combinations were tested intriplicate in adjacent wells on the same plate, and eachplate contained its own six untreated controls. All cock-tails and controls were within the 60 non-edge wells toavoid an edge effect.9,10 With each generation, weincluded 1�, 2�, and 4� single-agent controls in triplicatefor all 19 drugs on a separate plate to check forconsistency.
Preparation ofMultiple Cell LinesCells were isolated from BD Falcon 75-cm2 tissue
culture–treated flasks after a rinse with Cellgro PBS(Mediatech) and the addition of Gibco trypsin-EDTA(Invitrogen). The cells were plated at a volume of 50 AL/well in 96-well MicroWell Nunclon-D plates (Nalge NuncInternational). The numbers of cells per well were asfollows: A549, 1,500; H226, 3,000; PC-14, 2,500; HCC2279,4,000; HCC827, 7,000; H441, 2,700; H1975, 6,000; and Calu-3, 15,000. Cell lines were incubated on Nunc 167008 96-wellplates (Fisher Scientific International). After the drugmixture was added, the plates were incubated for 44 h at37jC in a humidified atmosphere containing 5% CO2. Thispermitted reproducible measurements and a repeat assayduring the workweek.
WST-1AssayCell proliferation was measured using the WST-1 Cell
Proliferation Reagent (Roche Diagnostics). After 44 h ofincubation, 25 AL of WST-1 reagent were added constitut-ing 9% of the well volume and the plates were incubatedfor another 4 h. The absorbance was measured on ascanning multiwell spectrophotometer at 440 nm with a600-nm reference.
Preparation of FSB/ASBSubsetsAll eight possible subcombinations of FSB—fenretinide,
suberoylanilide hydroxamic acid, and bortezomib—(F, S, B,FS, FB, SB, FSB, no treatment control) and ASB in whichfenretinide is replaced by all-trans retinoic acid (A, AS, AB,ASB) were tested in triplicate in eight NSCLC cell linesincluding A549. The doses were held stable in all cell linesat the IC10 for A549, and no edge wells were used. Eachplate contained six untreated controls. An additional plateof A549 cells containing 1�, 2�, and 4� doses in triplicateof each drug and six no treatment controls was done witheach experiment.
SynergyWe assessed the synergy of FSB in A549 and H2226 using
all 64 possible dose combinations derivable from four doselevels: zero, 1� (high), 0.66� (medium), and 0.44� (low),done in triplicate. We used the same doses in each cell lineto mirror the clinical case in which the drug titer inthe serum affects the tumor as a whole unvaried from oneclone to another. No edge wells were used and each platecontained six untreated controls. We also did single-agentcontrols on a separate plate at the 1� (high), 2�, and4� doses for both cell lines. These values were also used inthe synergy calculations. The combination index wascalculated by the Chou-Talalay equation. In cases in whichsingle agents stimulated growth, the value 0.000,001 wasused to allow a calculation.
Statistical AnalysesNo formal statistics were used to estimate the fitness.
Statistical analyses of the MACS data and proliferationassays were carried out in Excel (Microsoft, Inc.) and SPlus(Insightful, Inc.) using ANOVA. The Kolmogorov-Smirnovtest was run on the fitness values on sets of randomlyformed cocktails. Analyses of synergy were done usingCalcuSyn software (Biosoft Ltd.; refs. 16, 17).
Molecular Cancer Therapeutics 525
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
ResultsFitnessAn assessment of two sets of randomly formed cocktails
showed that each additional drug increased inhibition byf10%, supporting the 10% penalty for the fitness function(Supplementary Fig. S2).8
MACSThe hill climb-MACS arrived at the combination, FSB,
in only three generations (generations 0–2). It was
confirmed in the 4th generation (generation 3) as fittest
in its immediate neighborhood after assaying a total of
72 unique combinations over 2 weeks (Figs. 2A and 3D).
Figure 3. This chart shows the performance of all the cocktails in the hill climb across the four generations. With the exception of generation 0, thesecond column indicates the changes to the parent cocktail to form its nearest neighbors with each change resulting in a given ‘‘descendent.’’ The columnlabeled ‘‘FSB’’ indicates which among the three drugs are embedded in the cocktail. K(s) cellular activity compared with the no-treatment control. A samplefitness value calculation is provided using Bor,Ly,Sa, �(0.33 + 0.1(3)) = �0.66. The cocktails are sequenced by fitness. The fittest cocktail is highlightedand becomes parent to the next generation. If one of the progeny is fittest, a new color is introduced. A, cocktails in generation 0 were formed randomly. B,C and D, the parent and its 19 progeny (nearest neighbors). Abbreviations are derived from the first two or three letters of the name of each drug.
MACS to Identify Therapeutic Drug Combinations526
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
This is the same cocktail identified from MACSs described
elsewhere.9,10 We compared FSB and its nearest neighbors
to two separate sets of randomly formed cocktails:
18 cocktails forming generation 0 of the hill climb and
30 cocktails derived from earlier experiments.9,10 To
compare to the 18 cocktails, we used the lower of two
fitness values for FSB from generation 3 in the hill climb
[K(s) = 0.26; fitness, �0.56]. FSB was 2.57 SD above
the mean of its generation 0 (mean, �1.09; SD, 0.207)
or in the upper 0.51% of the distribution. We had
previously compared this same set (FSB and its nearest
neighbors) by testing it the same day as the 30 randomly
formed cocktails.9,10 FSB [K(s) = 0.24; fitness, �0.54]
performed 4.18 SD above the mean or in the upper
0.0014% of the distribution of these 30 cocktails assuming
a normal distribution of fitness levels (Fig. 2B).9,10 A
Kolmogorov-Smirnov test on multiple sets of randomly
formed cocktails derived from these earlier experiments
Figure 3 Continued .
Molecular Cancer Therapeutics 527
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
(including the 30 mentioned above) produced a P value of
0.5, indicating a close approximation of normality.In the hill climb, there was stability of the single-agent
controls across the generations (Fig. 2C; SupplementaryFig. S3)8 but a statistical improvement in the meanperformance of the latter generations compared with themean of the baseline generation (Figs. 2D and 3).Likewise, FSB and its nearest neighbors had a meanfitness value of �0.64 (SD, 0.07) compared with a meanfitness value of �1.18 (SD, 0.15) for the 30 randomlyformed cocktails (Fig. 2B). Thus, the hypothesis that themeans of the two distributions were the same was easilyrejected (t = 17.1).A closer look shows a steady enrichment for FSB.
Among all 18 randomly formed cocktails from generation0, the 6 most inhibitory cocktails contained a doublet
derived from FSB (Fig. 3). In generation 1, if the parent,Bor,Ly,Sa, had inhibited slightly more, it would have beenfittest and the hill climb would have stopped. Bygeneration 2, all the nearest neighbors of BS were testedand FSB was fittest. In generation 3, FSB as parent was thefittest in its immediate neighborhood and the hill climbstopped. Note that subtraction of any of the drugs from FSBincreased the K(s) sufficiently to render the doublets less fit.Moreover, addition of any drugs did not yield additionalimprovements in K(s) because FSB was maximally inhib-itory under the laboratory conditions we used.We compared FSB to 22 other triplets tested in
an earlier MACSs.9,10 The K(s) values were 0.43 to0.99 (fitness, �0.73 to �1.29), compared with 0.18 � 0.26(�0.48 to �0.56), among the six values for FSB in allMACSs.9,10 FSB was also superior to an additional
Figure 4. A, the synergy values and corresponding raw data for A549. B, the raw data for H226. The drugs and relative doses are indicated at theleft and above. A and B, yellow, single agents; blue, doublets; tan, triplets; purple, averages. SB dose combinations are shown twice and are indicatedusing the rounded rectangles. Single-agent doses used in combination and the additional doses used in the controls are listed together below. L, low;M, medium; H, high; H is the same dose as 1�; F, fenretinide; S, suberoylanilide hydroxamic acid; and B, bortezomib. Synergy is indicated by a value <1.For raw data, values are equal to the fraction; inhibition, respectively, and negative values indicating stimulation.
MACS to Identify Therapeutic Drug Combinations528
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
20 triplets examined through the hill climb with K(s)0.28 to 0.88 (�0.58 to �118; Fig. 3) although these tripletswere highly enriched for FSB.FSB was found to be synergistic in A549 (Fig. 4). In H226,
synergy could not be calculated for most dose combina-tions given the single-agent stimulatory activity of sub-eroylanilide hydroxamic acid despite assigning a value of0.000,001 to allow calculations (Fig. 4). This stimulatorybehavior confirms earlier dose response results (Fig. 5A)and experiments discussed below (Fig. 5B). Althoughsuberoylanilide hydroxamic acid increasingly stimulatedH226, as its doses increased as a single agent andin combination with fenretinide, it had minimal effectwhen paired with bortezomib and magnified inhibitionwhen combined with SB (Fig. 4).
Across all eight NSCLC cell lines, FSB was uniformlyinhibitory, whereas the doublets were not and the singleagents even less so (Fig. 5; Supplementary Table S1).8 Inall cases, FSB was more inhibitory than its constituentdoublets. The substitution of all-trans retinoic acid, acongener of fenretinide, when averaging across all eightcell lines, yielded less inhibitory doublets and tripletdespite the fact that all-trans retinoic acid, on average,was given at a more inhibitory single-agent dose thanfenretinide (Supplementary Table S1).8
DiscussionTo the best of our knowledge, this is the first study thatshows the feasibility for screening drug combinations of
Figure 5. The results from the power sets (i.e., all possible subsets) of FSB (FSB, FS, FB, SB, F, S, B, and no-treatment control) and of ASB (ASB, AS,AB, SB, A, S, B, and no-treatment control) across multiple NSCLC cell lines are shown. The y -axis shows the ratio of the WST-1 value of the treated overthe untreated control for a given cell line. Not shown are the A549 1�, 2�, and 4� single-agent controls that were done on the same day of allexperiments. The set of seven other cell lines were set up on 4 separate days. On each of these days, A549 was used a control. A, dose-response curvesfor suberoylanilide hydroxamic acid (SAHA ) in H226. The DMSO control titers match the increasing DMSO concentrations that increase as SAHA titersincrease. B, A549 and H226. H226 with the A549 control was tested on 2 separate days and both sets of results are shown. C, A549 and Calu-3. D,A549, H441, and H1975. E, A549, PC-14, HCC2279, and HCC827. The numerical values are shown in Supplementary Table S1.
Molecular Cancer Therapeutics 529
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
arbitrary size. It may also be the first using the strategy of asearch algorithm–guided screen of cocktails (i.e., a MACSas described earlier9,10). The hill climb-MACS quickly andautomatically identified a highly fit combination, FSB, thesame cocktail identified in an earlier MACS.9,10 It did sowithout the benefit of prior molecular insight into theinteractions of the combined drugs. In addition, use of anovel fitness function substantially increased efficiencies atthe level of the individual cocktail as compared withsynergy thereby facilitating MACS in our lab.Enrichment through the generations of sets of cocktails
and the superior fitness of the FSB itself indicate that thehill climb-MACS was operating effectively (Fig. 2). Thisenrichment is also represented by the progressivelyincreased presence of cocktails containing drugs derivedfrom FSB, which were more inhibitory, a behavior that alsoindicates the upward directedness of the screen (Fig. 3).Similar behavior was also observed in the earlierMACSs.9,10 Indeed, a MACS, whatever its rules, dependson the presence of a relationship between cocktailcomposition and fitness because the search has to ‘‘learn’’from prior experience.In the case of the hill climb-MACS, a triplet in the initial
generation was highly enriched for FSB, by chance,placing it relatively high up a peak on the fitness landscape(Fig. 3A). This may in part explain the rapidity withwhich the hill climb identified FSB as well as its relativelylow value of 2.57 SD above the mean fitness of itsgeneration 0. Indeed, of 969 possible triplets, only 4(0.41%), including FSB, contain at least two of the drugs.FSB at 4.18 SD above the mean needs to be interpreted withcaution because the random set of 30 cocktails and itscontrols included edge wells whereas FSB and its nearestneighbors and controls did not (despite both sets beingtested on the same day).9,10 A number of post hoc analysescontrolling for the edge effect yielded SD values rangingfrom 3.4 to >5 SD above the mean (data not shown). Evenat 3.4 SD, FSB is still higher than would have been expectedif cocktails were randomly tested (368 cocktails wereevaluated through the earlier MACSs). The interpretationof the significance of the number of SD also depends on theassumption of a normal distribution of fitness values.Although an analysis of a larger set of 60 randomly formedcocktails showed a close statistical approximation tonormality, we expect the tail of the distribution to ‘‘fatten’’as cocktails fully inhibit by being quite large. Nonetheless,it does seem that FSB performed better than expected basedon the above and other considerations.In addition, FSB was substantially fitter than any of the
other unrelated 22 triplets tested; thus, its fitness is not just asimple consequence of the fact that it is a triplet. Somecaution is warranted because some of these triplets and theirrespective controls were tested using edge wells whereasFSB and its controls were not. However, the edge effect issubstantially less than the observed differences in inhibitionbetween FSB and the other triplets (data not shown).The fact that FSB, which had been identified in an earlier
MACS,9,10 was identified a second time in the hill climb
after screening only 72 cocktails was highly unlikely bychance given the space of a total of >500,000 potentialcocktails and the many 10,000s of smaller potential cock-tails. This also suggests that this particular fitness land-scape is composed of a few higher peaks given theimprobability of ascending the same peak twice by chanceotherwise (Supplementary Fig. S1).8 Interestingly, the hillclimb nearly stopped at the triplet, Bor,LY,Sa, as itascended the peak. Indeed, if by chance it had been aslightly more inhibitory value, the hill climb would havestopped. This indicates the potential benefit in performinga number of hill climb-MACS runs even with simplerlandscapes.With complex multipeaked landscapes, a hill climb-
MACSs may arrive at different highly fit cocktails ondifferent runs, thereby supplying a number of variedcandidate cocktails (Supplementary Fig. S1).8 However,such fitness landscapes, especially when composed ofmany lesser peaks, are often more successfully canvassedby more complex search algorithm–guided screens such asthe hybrid algorithms we initially tested.9,10 Indeed, futureresearch can entail a determination of how factors such asassay type, constituent drugs, and varied fitness functionscan influence the topography of fitness landscapes.To speed the evaluation of individual cocktails, we used a
fitness function alternative to synergy during the MACS.However, synergy remains important to consider giventhat it is a mainstay of current cocktail assessments andarguably more reliably predicts clinical efficacy because itintegrates multiple dosing results per cocktail into asingle synergy value. However, synergy experimentsare increasingly labor-intensive as cocktails get largerwith exponential increases in dose permutations. More-over, such cocktails will likely have complex dose-responselandscapes as hinted at by our results (Fig. 4). Theselikely cannot be neatly summarized by a synergy value.In addition, stimulatory values are not accommodatedby standard synergy software, an issue that can ariseeven with single agents when using more than one cellline (Figs. 4 and 5A and B). One alternative to synergy toevaluate such dose-response combinatorial spaces is to usea search algorithm to guide a screen. This was recentlyreported by Wong et al. (15) on a space of 1,000,000 dosepermutations derived from six drugs at 10 different dosesper drug. Their strategy was similar to that used by us inthis article and as previously described9,10 but in their caseinstead directed at dose permutations. They used viralinhibition of a single-dose combination as a fitness valuerather than synergy and applied a stochastic searchalgorithm to arrive at a highly fit dose combination.However, one could examine dose-response fitness datato identify regions of the space simple enough to performstandard synergy analyses. Such spaces may be yet moreamenable to assessment using novel analytic synergymethods (18). Molecular insights may further directanalyses of this space.Although the method reported by Wong et al. was highly
efficient and reproducible, it required weeks to complete
MACS to Identify Therapeutic Drug Combinations530
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
a single run to arrive at the highly fit dose combination fora given cocktail. Their method applied to a large space ofpossible cocktails would thus be less speedy per cocktail,although more informative, than the fitness function wehad used to evaluate individual cocktails. Future study canexplore how to prioritize and optimize search effortsbetween the large space of different potential cocktails, aswe have done, and the space of dose permutations percocktail as done by Wong et al. (15). Yet other large combi-natorial spaces formed from drug sequencing and altereddose durations, among others, may also merit exploration.For MACS to be meaningful, the fitness result needs to be
predictive. Future study could evaluate this including ourfitness function, synergy, or any other function. Validationstudies showing activity of FSB in multiple cell lines suggestsome predictive potential of our simple fitness function.Moreover, it supports the hypothesis that combinationtherapy is more effective against a diversity of clones acrosspatients and within heterogeneous tumors, a plausibleobservation given the experience in the clinic. There are alsodata from other labs supporting the positive interactions ofsome of the constituent doublets of FSB, although FSB seemsto be a novel cocktail not previously studied (19–21). Thesereports indicate the power of existing and improvingmolecular insights in developing drug cocktails such asthe doublets mentioned above. MACS is intended tocomplement current discovery methods for combinationtherapy based on molecular insights thereby speeding theexisting process. It may also identify otherwise unantici-pated candidate cocktails. Indeed, the potential to findcocktails unexpected by available insights is also indicatedby results from a massive unguided screen of doublets byCombitoRx through which they found highly promisinganticancer doublets formed from noncancer drugs (10).The observation that there is a wide spectrum of NSCLC
lines inhibited by FSB suggests a potential to inhibit hostcells thereby drawing attention to the limitations of in vitroassays as predictors of toxicity. In the case of FSB, there aresome clinical data indicating that FSB may be tolerable inthe clinic (22–24). However, concerns for high risk oftoxicity of drug cocktails argue the interest in developmentof in vitro toxicity assays, which in turn could be integratedinto fitness functions and the MACSs. Indeed, the penaltywe affixed to larger cocktails in our fitness function was alimited attempt to represent toxicity by favoring smallercocktails. Whether FSB should be further developed ornot, although there are substantial predictive limitationsof existing in vitro assays, they are a foundation ofdrug development because they potentially offer someinformation of clinical value.Separate from the question of feasibility and predictive-
ness of a fitness function are barriers to exploring the entirespace of potential cocktails posed by limits in an assay. Forexample, we were unable to explore larger cocktailsembedded with FSB, which, like FSB, maximally inhibitedA549 but by definition were less fit. Alterations in assays orits conditions may expand the space of cocktails that can bemeaningfully compared.
Although F, S, and B are not dosed at high single-agentinhibitory values relative to the other agents, they dohave a disproportionately steep dose-response curve,with 4� dose inhibitory levels ranked 1st, 3rd, and 4th(all-trans retinoic acid was 16th; Table 1). MACS was notembedded with rules entailing prior knowledge of thisattribute and yet efficiently found FSB. Nonetheless, suchdrug activity information or other functional data obtainedthrough interrogating cocktail data sets generated during aMACS may usefully be embedded in a new algorithm orfitness function to improve MACS efficiencies.Likewise, future MACSs could integrate biological
insights by assigning a fitness ‘‘reward’’ for combinationsthat contain pairs of agents that are biologically tenable, oralternatively, one could choose pools of drugs that fit amolecular theme (or one could even use molecular ratherthan functional assays in the MACS). Contrariwise, MACSsfounded in functional assays can be used to identify notonly candidate therapeutics but also highly relevantcandidate molecular probes. Insights derived from theseprobes could in turn be integrated into a new MACS,thereby forming a positive feedback loop of alternatingMACSs and molecular evaluations. Limitations to molec-ular insights are exemplified by the fact that we would notlikely have studied a fenretinide-containing regimenwithout a MACS given the recent diminished clinicalinterest in retinoids in lung cancer. They are also illustratedby the knowledge that all-trans retinoic acid is a congenerof fenretinide and enabled only a limited capacity topredict efficacy (Supplementary Table S1).8 Thus, success-ful MACSs would likely not limit exploration to molecu-larly tenable cocktails.Further studies of combinatorial landscapes and search
algorithms as well as the development of efficient fitnessfunctions and improvements in laboratory gold standardsand clinical models to evaluate promising candidates needto be developed. Even if MACS is shown elsewhere to be auseful strategy, the need to conduct a lengthy series oflaboratory tests will remain, and if still promising, a seriesof evaluations through clinical trials of candidate cocktails.In addition, we note that there may be some challenges forwidespread adoption of MACS, particularly the contractualor legal difficulties that may present when testing cocktailswith agents owned by different entities. Principles derivedfrom the development of MACS in this oncology settingmay also be applied on a wider basis, such as modeling andassessing therapeutic candidates for infectious, rheumato-logic, or other diseases, providing additional impetus tofurther development of this potentially advantageousscreening strategy.
Disclosure of Potential Conflicts of InterestNo potential conflicts of interest were disclosed.
Acknowledgments
We thank Jean P. Issa, M.D., for kindly providing critical reagents; DafnaLotan, M.S., for her superb assistance in drug preparation; Frank Fossella,
Molecular Cancer Therapeutics 531
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
M.D., Mandri Obeyesekere, Ph.D., and Gabor Balazsi, Ph.D., for carefullyreading the text and their highly insightful editorial suggestions; andKathryn B. Carnes and Suzanne Davis, M.B.A., for excellent editorialassistance.
References
1. Sandler A, Gray R, Perry MC. Paclitaxel-carboplatin alone or withbevacizumab for non-small cell lung cancer. N Engl J Med 2006;355:2542–50.
2. Slamon DJ, Leyland-Jones B, Shak S, et al. Use of chemotherapyplus a monoclonal antibody against HER2 for metastatic breast cancerthat overexpresses HER2. N Engl J Med 2001;344:783–92.
3. Frei E III, Eder JP. Principles of dose, schedule, and combinationtherapy. In: Holland JF, Bast RC, Morton DL, Frei E, Kufe DW,Weichselbaum RR, editors. Cancer medicine. Baltimore: Williams andWilkins; 2006. p. 590–9.
4. Goldie JH, Codman AJ. A mathematical model for relating the drugsensitivity of tumors to the spontaneous mutation rate. Cancer Treat Rep1979;63:1727–33.
5. Chu E, DeVita VT, Jr. Principles of cancer management: chemother-apy. In: DeVita VT, Jr., Hellman S, Rosenberg SA, editors. Cancer,principles and practice of oncology. Philadelphia: Lippencott Williams &Wilkins; 2001. p. 289–306.
6. Bendell JC, George D, Nixon A, Yu D, Hurwitz H. Results of a phaseI study of bevacizumab (BV), everolimus (EV), and erlotinib (E) inpatients with advanced solid tumors. Proc Am Soc Clin Oncol2007;Abstract 3548.
7. Kwak EL, Clark JW, Chabner B. Targeted agents: the rules ofcombination. Clin Cancer Res 2007;13:5232–7.
8. Smith A. Screening for drug discovery: the leading question. Nature2002;418:453–9.
9. Goodnow RA, Jr., Gillespie P. Hit and Lead identification: efficientpractices for drug discovery. Prog Med Chem 2007;45:1–61.
10. Wadman M. The right combination. Nature 2006;439:390–401.
11. Holland JH. Adaptation in natural and artificial systems, 2nd ed. AnnArbor: MIT Press; 1992.
12. Kirkpatrick S, Gelatt CD, Jr., Vecchi MP. Optimization by SimulatedAnnealing. Science 1983;220:671–80.
13. Abbas HA. An evolutionary artificial neural networks approach forbreast cancer diagnosis. Artif Intell Med 2002;25:265–81.
14. Zheng B, Chang YH, Wang XH, Good WF. Feature selection forcomputerized mass detection in digitized mammograms by using a geneticalgorithm. Acad Radiol 1999;6:327–32.
15. Wong PK, Yu F, Shahangian A, Cheng G, Sun R, Ho CM. Closed-loopcontrol of cellular functions using combinatory drugs guided by astochastic search algorithm. Proc Natl Acad Sci U S A 2008;105:5105–10.
16. Chou TC, Talalay P. Quantitative analysis of dose-effect relationships:the combined effects of multiple drugs or enzyme inhibitors. Adv EnzRegul 1984;22:27–55.
17. Chou TC, Motzer RJ, Tong Y, Bosl GJ. Computerized quantitation ofsynergism and antagonism of Taxol, topotecan and cisplatin againstteratocarcinoma cell growth: a rational approach to clinical protocoldesign. J Natl Cancer Inst 1994;86:1517–24.
18. Fang HB, Ross DD, Sausville E, Tan M. Experimental design andinteraction analysis of combination studies of drugs with log-linear doseresponses. Stat Med 2008;27:3071–83.
19. Whang YM, Choi EJ, Seo JH. Hyperacetylation enhances the growth-inhibitory effect of all-trans retinoic acid by the restoration of retinoic acidreceptor h expression in head and neck squamous carcinoma (HNSCC)cells. Cancer Chemother Phamacol 2005;56:543–55.
20. Denlinger CE, Rundall BK, Jones DR. Proteasome inhibition sensitizesnon-small cell lung cancer to histone deacetylase inhibitor-inducedapoptosis through the generation of reactive oxygen species. J ThoracCardiovasc Surg 2004;128:740–8.
21. Touma SE, Goldberg JS, Moench P. Retinoic acid and the histonedeacetylase inhibitor trichostatin a inhibit the proliferation of human renalcell carcinoma in a xenograft tumor model. Clin Cancer Res 2005;11:3558–66.
22. Schelman WR, Kolesar J, Schell K, et al. A phase I study of vorinostatin combination with bortezomib in refractory solid tumors. Journal ofClinical Oncology, 2007 ASCO Annual Meeting Proceedings Part I. Vol 25,No. 18S, 2007:3573.
23. Badros AZ, Philip S, Niesvizk R, et al. Grant Phase I trial of vorinostatplus bortezomib (bort) in relapsed/refractory multiple myeloma (mm)patients (pts). J Clin Oncol 2008;26:abstr 8548.
24. Costa A, Malone W, Perloff M, et al. Tolerability of the syntheticretinoid fenretinide (HPR). Eur J Cancer Clin Oncol 1989;25:805–8.
MACS to Identify Therapeutic Drug Combinations532
Mol Cancer Ther 2009;8(3). March 2009
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937
Published OnlineFirst March 10, 2009.Mol Cancer Ther arbitrary size for activity against cancer cellsAlgorithmic guided screening of drug combinations of
Updated version
10.1158/1535-7163.MCT-08-0937doi:
Access the most recent version of this article at:
Material
Supplementary
http://mct.aacrjournals.org/content/suppl/2009/03/04/1535-7163.MCT-08-0937.DC1
Access the most recent supplemental material at:
E-mail alerts related to this article or journal.Sign up to receive free email-alerts
Subscriptions
Reprints and
To order reprints of this article or to subscribe to the journal, contact the AACR Publications
Permissions
Rightslink site. (CCC)Click on "Request Permissions" which will take you to the Copyright Clearance Center's
.http://mct.aacrjournals.org/content/early/2005/11/01/1535-7163.MCT-08-0937.citationTo request permission to re-use all or part of this article, use this link
Research. on August 14, 2021. © 2009 American Association for Cancermct.aacrjournals.org Downloaded from
Published OnlineFirst March 10, 2009; DOI: 10.1158/1535-7163.MCT-08-0937